About the Role
You'll own the pipeline from teleoperation data → trained models → deployed inference. This is the "make robots learn from humans" role. We're looking for someone who's played with lerobot, OpenTeleVision, PI0, or similar VLA/imitation learning codebases and can ship production systems—not just train models in notebooks.
What You'll Do
- Build data collection pipelines that capture teleoperation sessions (RGB, depth, joint states, force/torque, gripper position) at 50-250Hz
- Create annotation and quality-checking tools for teleoperation data
- Set up cloud training infrastructure for fine-tuning manipulation models
- Optimize inference for edge deployment on robot controllers
- Work closely with the teleoperation engineer to define what data to collect and how
- Build ETL systems that process terabytes of robot sensor data into training-ready datasets
What We're Looking For
- Proficiency in Python and at least one deep learning library (PyTorch, JAX, or TensorFlow)
- CUDA experience—you've optimized training or inference, not just called .cuda()
- Familiar with open-source robotics ML: lerobot, OpenTeleVision, PI0, or similar VLA/imitation learning codebases
- Deep understanding of state-of-the-art ML techniques for robotics (behavior cloning, diffusion policies, transformer architectures for control)
- Experience with cloud-based training environments (AWS, GCP, or Azure)
- Familiarity with simulation tools (Isaac Sim, Gazebo, PyBullet, MuJoCo)
- Experience deploying models on embedded hardware (Jetson or similar)
- Comfortable working in Linux-based environments
- Has deployed something in production, not just trained models in notebooks
Why Join Us
We make robot deployment work like SaaS—browse any robot, deploy with one click. Fresh off a contract with KUKA Robotics deploying our software across the world's largest appliance manufacturer. Two founders, no fluff. You'll join early and own the entire AI and data stack.